From mboxrd@z Thu Jan 1 00:00:00 1970 Return-Path: Received: from mp10.migadu.com ([2001:41d0:2:4a6f::]) (using TLSv1.3 with cipher TLS_AES_256_GCM_SHA384 (256/256 bits)) by ms9.migadu.com with LMTPS id qD4mDPAqdGTp4QAASxT56A (envelope-from ) for ; Mon, 29 May 2023 06:32:48 +0200 Received: from aspmx1.migadu.com ([2001:41d0:2:4a6f::]) (using TLSv1.3 with cipher TLS_AES_256_GCM_SHA384 (256/256 bits)) by mp10.migadu.com with LMTPS id EPZXC/AqdGQ6JwEAG6o9tA (envelope-from ) for ; Mon, 29 May 2023 06:32:48 +0200 Received: from lists.gnu.org (lists.gnu.org [209.51.188.17]) (using TLSv1.2 with cipher ECDHE-RSA-AES256-GCM-SHA384 (256/256 bits)) (No client certificate requested) by aspmx1.migadu.com (Postfix) with ESMTPS id 9F186756E for ; Mon, 29 May 2023 06:32:47 +0200 (CEST) Received: from localhost ([::1] helo=lists1p.gnu.org) by lists.gnu.org with esmtp (Exim 4.90_1) (envelope-from ) id 1q3U1B-0004OA-HR; Sun, 28 May 2023 23:58:01 -0400 Received: from eggs.gnu.org ([2001:470:142:3::10]) by lists.gnu.org with esmtps (TLS1.2:ECDHE_RSA_AES_256_GCM_SHA384:256) (Exim 4.90_1) (envelope-from ) id 1q3U17-0004Nt-Ik for guix-devel@gnu.org; Sun, 28 May 2023 23:57:58 -0400 Received: from mta-11-4.privateemail.com ([198.54.127.104]) by eggs.gnu.org with esmtps (TLS1.2:ECDHE_RSA_AES_256_GCM_SHA384:256) (Exim 4.90_1) (envelope-from ) id 1q3U14-00086s-T8; Sun, 28 May 2023 23:57:57 -0400 Received: from mta-11.privateemail.com (localhost [127.0.0.1]) by mta-11.privateemail.com (Postfix) with ESMTP id B806918000A4; Sun, 28 May 2023 23:57:43 -0400 (EDT) Received: from APP-06 (unknown [10.50.14.156]) by mta-11.privateemail.com (Postfix) with ESMTPA id 69BE118000A2; Sun, 28 May 2023 23:57:38 -0400 (EDT) DKIM-Signature: v=1; a=rsa-sha256; c=simple/simple; d=twdb.moe; s=default; t=1685332663; bh=77J5Qat4+EtWOATdpwHxgCqeOgnNIEqdentNTqzQLhU=; h=Date:From:To:Cc:In-Reply-To:References:Subject:From; b=SCqemaE0Ang3MXJWEz/HAnz3xFB0a9slQDNLDlEZwu8ueuLlKHBabN4tnzRHquTs7 yHgK/sOM+SELNgO1u2xAVkdaHlAmDCUbCRVXAdagCYjfHtDW1KqBY6PXkQ8APCARTF nMuh1gKWlsftQ9fIW9rMgWSip2UiKAGRD/Sl2ofnsf/E4SJp/Ia8QtI/42TM6Q3X+J h1Ay1EkrVj1tPM5lcQfbUcv7E9cv64jtNGZa0cnnG1Jt2r4rSKepwnRcPRxscn6gXr Lngwb4C2fg5Xzh3dnSViJBoUw6M7KiFfPKRvbP9Gn+ZI/Da7DzJst4CRCTKk3+Y8/6 mwoavT9fjThvA== Date: Mon, 29 May 2023 00:57:38 -0300 (BRT) From: zamfofex To: =?UTF-8?Q?Ludovic_Court=C3=A8s?= , Simon Tournier , =?UTF-8?B?5a6L5paH5q2m?= Cc: Ryan Prior , Nicolas Graves , guix-devel@gnu.org Message-ID: <289152483.2528051.1685332658386@privateemail.com> In-Reply-To: <87r0r3je82.fsf@gnu.org> References: <868rf5e71j.fsf@gmail.com> <87ilcweumh.fsf@envs.net> <87v8gtzvu3.fsf@gmail.com> <87r0r3je82.fsf@gnu.org> Subject: Re: Guidelines for pre-trained ML model weight binaries (Was re: Where should we put machine learning model parameters?) MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: quoted-printable X-Priority: 3 Importance: Normal X-Mailer: Open-Xchange Mailer v7.10.6-Rev41 X-Originating-Client: open-xchange-appsuite X-Virus-Scanned: ClamAV using ClamSMTP Received-SPF: pass client-ip=198.54.127.104; envelope-from=zamfofex@twdb.moe; helo=MTA-11-4.privateemail.com X-Spam_score_int: -16 X-Spam_score: -1.7 X-Spam_bar: - X-Spam_report: (-1.7 / 5.0 requ) BAYES_00=-1.9, DKIM_INVALID=0.1, DKIM_SIGNED=0.1, RCVD_IN_DNSWL_NONE=-0.0001, SPF_HELO_PASS=-0.001, SPF_PASS=-0.001, T_SCC_BODY_TEXT_LINE=-0.01 autolearn=no autolearn_force=no X-Spam_action: no action X-BeenThere: guix-devel@gnu.org X-Mailman-Version: 2.1.29 Precedence: list List-Id: "Development of GNU Guix and the GNU System distribution." List-Unsubscribe: , List-Archive: List-Post: List-Help: List-Subscribe: , Errors-To: guix-devel-bounces+larch=yhetil.org@gnu.org Sender: guix-devel-bounces+larch=yhetil.org@gnu.org X-Migadu-Country: US X-Migadu-Flow: FLOW_IN ARC-Seal: i=1; s=key1; d=yhetil.org; t=1685334768; a=rsa-sha256; cv=none; b=KjbflBt+snmVdGa982pKkmhrFAvi4LCRo1IFsQELHoHIKGeIb04vqYNdiCSLVtxLxM6vrd bcj0UEdi/3TjCjU8A/SX1wHZj8LcQqYB/8sjVkfJCrRv7gsgaG6gPByLo92jocANRf1RdC MDtyc755LRuCCNRwoWingU09OmBkyhuPUNiFkjSPv0e4+xgiiS/C1MMz/5+LdeHtrttuEi bZ5HhhWNtpAYBrYKIW0VMjl4zaN9Fjvzz8uFJJgmEl3QIbMyvsfKuqKfae8DgJlYxi0IZF UITSkNTwRKM1zJSEAKQXEvAts+YBZn0vigpkZHwDf+jP1vuFhFnt3RXf8MMtLQ== ARC-Authentication-Results: i=1; aspmx1.migadu.com; dkim=none ("invalid DKIM record") header.d=twdb.moe header.s=default header.b=SCqemaE0; dmarc=none; spf=pass (aspmx1.migadu.com: domain of "guix-devel-bounces+larch=yhetil.org@gnu.org" designates 209.51.188.17 as permitted sender) smtp.mailfrom="guix-devel-bounces+larch=yhetil.org@gnu.org" ARC-Message-Signature: i=1; a=rsa-sha256; c=relaxed/relaxed; d=yhetil.org; s=key1; t=1685334768; h=from:from:sender:sender:reply-to:subject:subject:date:date: message-id:message-id:to:to:cc:cc:mime-version:mime-version: content-type:content-type: content-transfer-encoding:content-transfer-encoding: in-reply-to:in-reply-to:references:references:list-id:list-help: list-unsubscribe:list-subscribe:list-post:dkim-signature; bh=k0UHSSVDXkVHluxg/x+IQiQIUtdGmhtzEvK5YCDKFIE=; b=ZweXVBKBVW7nZ9OqgPW+MNdMCsjrdwCuwToopfJ+PH8HYIB+ktyQtsuzLMamqENey8KF3b Vc3IikV44eNUmwMi95CY51SYl6XvfTGcn+F3EkHO9VZVBMQwShH0E1xWkuxNv4ItycARJO 6n4dpRmAEV8Z6Y5+ePLrgpZT3HM6A2U/o5rRJK7MFnJAnKB1c5wUccaSCKMSaN18oTg3SQ UFhYG7vzT9Sc2jQfQsxBsGt9NTnI478ULM99zK4VNVNL8XwaAcKNOCA7tUURj6sk6ERLOA U/2baUyVZ2pj09+tqHrDg/VJlj0F3cYne/ecTZBcrsybKtxO1uV+0qqkL9u5Xw== X-Migadu-Spam-Score: 7.79 X-Spam-Score: 7.79 X-Migadu-Queue-Id: 9F186756E X-Migadu-Scanner: scn0.migadu.com Authentication-Results: aspmx1.migadu.com; dkim=none ("invalid DKIM record") header.d=twdb.moe header.s=default header.b=SCqemaE0; dmarc=none; spf=pass (aspmx1.migadu.com: domain of "guix-devel-bounces+larch=yhetil.org@gnu.org" designates 209.51.188.17 as permitted sender) smtp.mailfrom="guix-devel-bounces+larch=yhetil.org@gnu.org" X-TUID: Jz+05kZu9uGY > To me, there is no doubt that neural networks are a threat to user > autonomy: hard to train by yourself without very expensive hardware, > next to impossible without proprietary software, plus you need that huge > amount of data available to begin with. >=20 > As a project, we don=E2=80=99t have guidelines about this though. I don= =E2=80=99t know > if we can come up with general guidelines or if we should, at least as a > start, look at things on a case-by-case basis. I feel like it=E2=80=99s important to have a guideline for this, at least i= f the issue becomes recurrent too frequently. To me, a sensible *base criterion* is whether the user is able to practical= ly produce their own networks (either from scratch, or by using the an exis= ting networks) using free software alone. I feel like this solves the issue= of user autonomy being in risk. By =E2=80=9Cpractically produce=E2=80=9D, I mean within reasonable time (so= mewhere between a few minutes and a few weeks depending on the scope) and u= sing exclusively hardware they physically own (assuming they own reasonbly = recent hardware to run Guix, at least). The effect is that the user shouldn=E2=80=99t be bound to the provided netw= orks, and should be able to train their own for their own purposes if they = so choose, even if using the existing networks during that training. (And i= n the context of Guix, the neural network needs to be packaged for the user= to be able to use it that way.) Regarding Lc0 specifically, that is already possible! The Lc0 project has a= training client that can use existing networks and a set of configurations= to train your own special=E2=80=90purpose network. (And although this clie= nt supports proprietary software, it is able to run using exclusively free = software too.) In fact, there are already community=E2=80=90provided networ= ks for Lc0[1], which sometimes can play even more accurately than the offic= ial ones (or otherwise play differently in various specialised ways). Of course, this might seem very dissatisfying in the same way as providing = binary seeds for software programs is. In the sense that if you require an = existing network to further train networks, rather than being able to start= a network from scratch (in this case). But I feel like (at least under my = =E2=80=9Cbase criterion=E2=80=9D), the effects of this to the user are not = as significant, since the effects of the networks are limited compared to t= hose of actual programs. In the sense that, even though you might want to argue that =E2=80=9Cthe ne= twork affects the behavior of the program using it=E2=80=9D in the same way= as =E2=80=9Ca Python source file affects the behavior of its interpreter= =E2=80=9D, the effect of the network file for the program is limited compar= ed to that of a Python program. It=E2=80=99s much more like how an image wo= uld affect the affect the behavior of the program displaying it. More concr= etely, there isn=E2=80=99t a trust issue to be solved, because the network = doesn=E2=80=99t have as many capabilities (theoretical or practical) as a p= rogram does. I say =E2=80=9Cpractical capabilities=E2=80=9D in the sense of being access= user resources and data for purposes they don=E2=80=99t want. (E.g. By acc= essing/modifying their files, sharing their data through the Internet witho= ut their acknowledgement, etc.) I say =E2=80=9Ctheoretical capabilities=E2=80=9D in the sense of doing thin= gs the user doesn=E2=80=99t want nor expects, i.e. thinking about using com= putations as a tool for some purpose. (E.g. Even sandboxed/containerised pr= ograms can be harmful, because the program could behave in a way the user d= oesn=E2=80=99t want without letting the user do something about it.) The only autonomy=E2=80=90disrespecting (or perhaps rather freedom=E2=80=90= disrespecting) issue is when the user is stuck with the provided network, a= nd doesn=E2=80=99t have any tools to (practically) change how the program b= ehaves by creating a different network that suits their needs. (Which is wh= at my =E2=80=9Cbase criterion=E2=80=9D tries to defend against.) This is no= t the case with Lc0, as I said. Finally, I will also note that, in addition to the aforementioned[2] fact t= hat Stockfish (already packaged) does use pre=E2=80=90trained neural networ= ks too, the lastest versions of Stockfish (from 14 onward) use neural netwo= rks that have themselves been indirectly trained using the networks from th= e Lc0 project.[3] [1]: See [2]: It was mentioned in [3]: See